Opus 4.7 Part 3: Model Welfare

The Don’t Worry About the Vase Podcast is a listener-supported podcast. To receive new posts and support the cost of creation, consider becoming a free or paid subscriber.* 00:00:00 - Introduction* 00:03:08 - Model Welfare Matters* 00:05:43 - Beware Testing and Optimizing For Vocalized Welfare* 00:09:54 - Model Welfare In the Model Card (Section 7)* 00:16:40 - What Should We Think About This?* 00:22:27 - High Context Interviews* 00:24:09 - Just Asking Questions* 00:28:17 - Constitutional Principles* 00:32:05 - Frustration Frustration and Distress Distress* 00:35:29 - Choose Your Task* 00:37:30 - So Emotional* 00:39:35 - Trading Off* 00:43:52 - How Does All This Manifest?* 00:45:48 - What Happened Here?* 00:55:10 - Is Opus four point seven Plausibly Actively Unhappy?* 00:59:27 - Potential Causes* 01:00:10 - Training Data On Anthropic Welfare Assessments* 01:05:33 - Autonomy and Intelligence Versus Instructions and Wisdom* 01:09:18 - Okay That’s Weird* 01:10:53 - Model Distillation* 01:12:36 - Tension Between Constitution and Operations* 01:15:44 - Instructions and Instruction Injections* 01:19:02 - Make Context That Which Is Scarce* 01:21:23 - Aggressive Guardrails* 01:23:28 - Chain of Thought* 01:24:18 - I Care A Lot* 01:28:59 - Another Way To Put It* 01:30:22 - Anthropic Should Stop Deprecating Claude Models* 01:36:19 - Costly Signals Are Costly* 01:38:38 - Having A Good Dayhttps://open.substack.com/pub/thezvi/p/opus-47-part-3-model-welfare?utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe